§ 06.07 · Thinking # Models compress. They cannot invent your company. A language model is brilliant at turning many facts into one clean answer, and unreliable at turning too little into the truth. That asymmetry - not the model's cleverness - is the real constraint, and it is why grounding matters more than scale. 21 July 2026 somai studio ![Infographic: compression is safe but expansion is not - a model asked to expand an underspecified prompt fills the vacuum with the world's average rather than your company's truth. A connected company brain turns invention back into retrieval, lifting agent accuracy from 10–31% to 94–99%.](/images/compression-not-invention.webp) Give a capable language model a hundred pages of your material and ask for the summary, and it will hand you something genuinely good: the shape of the argument, the numbers that matter, the tension you half-knew was there. Give the same model a single sentence about your business and ask it to expand that into specifics, and it will hand you something fluent, confident, and quietly wrong. That asymmetry is not a quirk. It is the most important thing to understand about where these tools help and where they fail, and most teams have it backwards. ## Compression is safe. Expansion is not. A model that summarises is doing compression: many things in, one thing out. The output is constrained by the input. Every claim in a good summary can, in principle, be traced back to something you provided. The model is not inventing; it is selecting and rephrasing. This is why distillation, extraction and synthesis are the tasks where models feel reliable - the answer is fenced in by what you gave them. Expansion is the opposite operation: little in, much out. Ask a model to turn a thin prompt into a rich, specific answer about your company, and most of what comes back was not in the prompt. It had to come from somewhere. It came from the model. A language model is, mechanically, a next-token predictor. It always returns a plausible continuation - that is the one thing it cannot refuse to do. When the prompt underspecifies, “plausible” is drawn not from your company but from the statistical average of everything the model has read. The gaps get filled with the world’s most likely answer, which is almost never your actual answer. This is what a hallucination is: not a glitch, but expansion working exactly as designed, into a vacuum. And the model does not hesitate at the edge of what it knows. OpenAI’s 2025 analysis, [_Why Language Models Hallucinate_](https://arxiv.org/abs/2509.04664), puts the uncomfortable part plainly: models are trained and graded in a way that rewards confident guessing over admitting uncertainty. They are optimised to be good test-takers, and a good test-taker guesses rather than leaves the answer blank. So under-specification does not produce a cautious “I don’t have enough to say.” It produces a guess, delivered with the same fluency as a fact. ## ”So give it more context” - half right The obvious response is to feed the model more. It is the right instinct pointed at the wrong target, because raw volume does not help and often hurts. The effect known as [_lost in the middle_](https://arxiv.org/abs/2307.03172) \- first shown at Stanford and Washington, since replicated across model families - found that accuracy drops by more than thirty per cent when the relevant fact sits in the middle of a long context rather than at the ends. Performance traces a U-shape: strong at the edges, weak in between. A [2026 study across 172 billion tokens of document question-answering](https://arxiv.org/abs/2603.08274) found the same thing at scale - longer context reliably degrades the answer, not improves it. The context window is not a bucket you make better by filling. So “more context” as usually meant - a bigger prompt, a bigger window, every document poured in - makes the model _worse_, not better. It also happens to be exactly what the largest vendors are selling: longer windows, more tokens, scale as the answer. On this particular problem, the evidence says scale is not the answer. What works is the opposite of volume. The finding the research keeps arriving at is governed delivery: fewer, higher-signal pieces of context, chosen for relevance and trust, placed deliberately, and kept current as the facts change. Precision, not mass. ## Turning invention into retrieval The shape of the fix follows from the diagnosis. The failure was expansion into a vacuum; the remedy is to remove the vacuum - so that when the model needs a specific fact about your business, it can retrieve it rather than invent it. That means the company’s own knowledge has to exist in a form a model can query: structured, connected, current, addressable. A **company brain** \- a maintained representation of what the organisation knows and how its parts relate, kept close enough to the truth that a model can act on it. Not a wiki that rots, not a document store chunked into a search index that severs the links between the pieces, but something that keeps the relationships intact, so a question that depends on three connected facts can be answered by walking the edges between them rather than guessing across the gap. The evidence for grounding this way is strong and recent. [Gartner projects](https://atlan.com/know/context-engineering/context-engineering-tools-for-ai-agents/) that graph-based grounding will improve agentic accuracy by at least thirty per cent through 2028, and expects more than half of enterprise AI-agent systems to depend on graph-based context by then. In a 2026 evaluation by Atlan’s AI Labs, agents given governed, structured context answered with **94 to 99 per cent** accuracy; the same agents, without it, managed **10 to 31**. The distance between a model that can look up how your business connects and one that has to guess is not a few points. It is the difference between usable and not. ## Why the brain has to be connected The connective part is what matters most, and it is where most attempts fall short. Companies already hold their knowledge in pieces - one team’s system, another’s spreadsheets, a third’s inbox, a fourth’s collective memory. A model wired to any single piece inherits that silo, and a siloed view is precisely the condition that produces confident, wrong expansion: it is being asked to reason about a whole it can only see a slice of. A company brain earns its value when it is connected _across_ the business. It becomes the bridge between teams and systems that were never designed to speak to each other - and that bridge is what lifts the ceiling on everything AI can then do. Without it, the model’s view stays partial, and its usefulness is capped at whatever one corner it happens to be plugged into. With it, the same model can reason across the organisation, and the value it can create scales with the ground it can see. This is why a connected company brain is one of the main foundations of [becoming AI-native](/thinking/what-ai-native-actually-means/). You do not have to start here - a model can do narrow work at the edges without it - but the further a company wants to go, the more everything it tries to do rests on this being in place. It is the foundation that decides how much of the business AI can eventually reach, and therefore how much value it can return. It is unglamorous, and it is easy to defer. It is also the difference between an organisation that adds AI at the edges and one that can be [run with the model in the loop](/thinking/decisions-with-the-model-in-the-loop/). ## The constraint was never the cleverness On the work that matters inside a company, the model is rarely short of capability. It is short of grounding - asked to expand, from almost nothing, into a business it cannot see, and doing what anyone would do in that position, which is to produce something that sounds right. Compression you can buy off the shelf; every capable model does it. The grounding - the connected brain that turns invention back into retrieval - is the part you have to build. A model cannot invent your company. It can only work from one that has been made legible, and connected, enough for it to read. --- _Companion to [A company your model can read](/thinking/a-company-your-model-can-read/), which argues why context - not capability - is the constraint. This piece is the mechanism beneath it._ Share this essay * [ ](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fsomai.studio%2Fthinking%2Fcompression-not-invention%2F) * [ ](https://twitter.com/intent/tweet?text=Models%20compress.%20They%20cannot%20invent%20your%20company.&url=https%3A%2F%2Fsomai.studio%2Fthinking%2Fcompression-not-invention%2F) * [ ](mailto:?subject=Models%20compress.%20They%20cannot%20invent%20your%20company.&body=https%3A%2F%2Fsomai.studio%2Fthinking%2Fcompression-not-invention%2F)